Power transmission lines (TLs) often lead to large-scale power outages caused by insulator failures, which not only disrupt normal production and daily life but also result in considerable economic losses. To enable the recognition of multiple defects in transmission line insulators, this paper proposes a deep learning method called coordinate attention-you only look once (CA‐YOLO). To address the issue of insufficient feature extraction in the original network, a convolutional block attention module (CBAM) is added after each C3 module in the backbone network, enhancing the feature extraction capability of the backbone with minimal computational overhead. To address the issue of effective information loss caused by scale inconsistency during feature fusion, the feature fusion architecture of YOLO has been restructured based on an adaptive spatial feature fusion (ASFF) network. This approach better manages the importance of feature maps by learning fusion features across different levels. An enhanced mosaic method is proposed, which generates new images by randomly cropping, scaling and arranging images, thereby enriching the dataset and improving the generalisation performance of the model. Experiments conducted on the transmission line insulator defect dataset constructed for this study show that the CA‐YOLO model achieves a 5.9% improvement in mean average precision (mAP) compared with the baseline model, with a detection speed of 76.9 fps, fulfilling the requirements for practical engineering applications.
Zhao et al. (Thu,) studied this question.